GPT-Live: OpenAI's Real-Time Voice AI, Explained Simply

OpenAI's GPT-Live now powers ChatGPT Voice, and it works differently from the voice mode before it: it can listen and speak at the same time, and it hands hard questions to GPT-5.5 while the conversation keeps going. We explain what changed, who gets which version, and what it costs developers.

What “full-duplex” means

ChatGPT's older Advanced Voice Mode worked in turns: you spoke, it waited for you to finish, then it replied. GPT-Live is full-duplex, meaning it listens and speaks at the same time. It can say “mhmm” while you are still talking, stay quiet when you pause to think, and stop when you cut in, which is closer to how two people actually talk. OpenAI released GPT-Live-1 and GPT-Live-1 mini on July 8, 2026, as reported by DeepLearning.AI's The Batch.

One model talks, another thinks

A voice model that has to answer instantly has little time to reason. GPT-Live solves this by splitting the work. The voice model handles the conversation, and when a question needs web search, deeper reasoning or several tool steps, it hands the job to GPT-5.5 and keeps talking while that runs, then works the answer in.

On GPT-Live-1 you can pick the reasoning effort. Instant uses GPT-5.5 Instant in the background; Medium and High use GPT-5.5 Thinking. GPT-Live-1 mini only calls GPT-5.5 Instant.

Who gets which version

GPT-Live-1 is the default voice on ChatGPT Go, Plus and Pro at no extra charge, and GPT-Live-1 mini is the default on the Free plan. It runs in ChatGPT Voice on iOS, Android and ChatGPT.com. At launch, OpenAI said it was still working on live video and screen sharing, which the old voice mode offered.

How good is it?

OpenAI's own tests compare GPT-Live only with the voice mode it replaced, not with rival voice assistants. In those tests, human raters preferred GPT-Live-1 over Advanced Voice Mode 75.7% of the time (69.2% for the mini model). With high reasoning, GPT-Live-1 scored 84.2% on the GPQA science benchmark, against 45.3% for the old voice mode. OpenAI has not published latency figures, so treat speed claims you see elsewhere with care.

GPT-Live for developers

GPT-Live-1 is now available in the API through a dedicated Live sessions endpoint. According to OpenAI's model page, voice sessions cost $0.05 per minute, billed per second, and the backend model and tools are billed separately at their normal rates. It accepts audio and text, supports function calling, and does not take images or video. Voice platforms such as LiveKit already support it.

Our take: what a 10-minute support call really costs

The $0.05 per minute headline is only part of the bill, because the thinking is charged separately. Here is a rough way to estimate it for a customer-support voice agent:

• Voice time: a 10-minute call is 10 × $0.05 = $0.50 in session fees.
• Backend model: every hand-off to GPT-5.5 adds normal token charges. A call that needs a few order look-ups costs more than one that is pure small talk.
• Reasoning effort: Instant keeps backend costs and waiting time down; save Medium or High for calls that genuinely need multi-step reasoning.
• Silence still counts: billing is per second of session, so end sessions promptly when the caller goes quiet or hangs up.

For everyday ChatGPT users the trade-off is simpler: GPT-Live is included in your plan, it feels much more natural to interrupt, and hard questions take a moment longer because they are being passed to a stronger model. As eesel AI's explainer notes, that split is the main design idea behind it.

Sources

• DeepLearning.AI, The Batch: GPT-Live pairs full-duplex voice models with GPT-5.5
• OpenAI developer docs: GPT-Live-1 model page
• eesel AI: GPT-Live-1, OpenAI's full-duplex voice model, explained
• LiveKit docs: GPT-Live realtime model

Frequently asked questions

Related articles

Comments